Netty: perf tests

[kciesielski]

Recently our Netty backends have fully switched to reactive-streams powered processing (see #3337).
Even raw requests and responses are now represented as Subscribers and Publishers, which had to be implemented manually for Future and Loom backends. The implementation is rather naive:

Request Body

• FileWriterSubscriber requests exactly 1 next chunk after writing current chunk.
• Similarly for InputStreamBody, where we load data using our SimpleSubscriber before exposing it to the logic as an InputStream.

Response Body

• FileRangePublisher reads in chunks and manually serializes requests using an AtomicBoolean
• InputStreamPublisher works similarly, it also uses a repeated blocking {} call to read from an InputStream

We'd like to measure how fast are these implementations, possibly compared to the previous impl (1.9.1) and streaming-based Netty backends like cats and zio, which work in a different way:

• Publisher from the request is converted into a fs2.Stream or ZStream using library code
• The obtained stream is then written to a file, also using library code

This issue has been created as a bookmark after noticing that our new backends elements may impact performance. Let's start with specifying what exactly we want to measure and compare.

[kciesielski]

https://softwaremill.com/benchmarking-tapir-part-1/